Key Takeaways
- OpenAI introduces Private Safety Processing, boosting safety without exposing underlying content, while maintaining zero data retention for frontier models.
- Codex demonstrates extreme automation: Asana completes a five-year frontend test migration in two weeks, and a tax-prep pilot cuts time by a third.
- Stripe acquires OpenRouter, though the price is disputed (undisclosed vs. $7 billion), highlighting model routing as a new AI infrastructure battleground.
- NVIDIA Vera Rubin delivers 10x more tokens per megawatt on CoreWeave and is already powering OpenAI training; cuOpt and DGX performance specs also debut.
- Hamel Husain and Ethan Mollick weigh in on AI writing quality and skill-building tools, while Perplexity launches Computer in Email for document revision.
1. OpenAI Enterprise Privacy and Safety
- OpenAI announced Private Safety Processing, a mechanism designed to improve safety without allowing OpenAI personnel to access underlying content, and reaffirmed zero data retention for frontier models. CEO Sam Altman also voiced support for enterprise privacy. This addresses core data-leakage concerns, potentially accelerating enterprise adoption of AI. — via 1 2 3
2. AI Programming and Automation: Codex
- Codex processed 7,000 filings in a tax-preparation pilot, saving about one-third of the time; developers can also build their own products on the open-source Codex harness. This shows Codex expanding from coding assistance to broad business automation. — via 1
- Codex compresses code migration from years to weeks: Asana used it to finish a frontend test migration in two weeks that was expected to take five years. The case is cited by both Greg Brockman and swyx, making it a benchmark for AI-driven engineering velocity. — via 1 2
- Codex can also handle a wide range of browser and desktop automation tasks, including accounting close, email triage, Stripe rule configuration, security questionnaires, and subscription cancellations—claiming to handle almost any browser task. — via 1
3. AI Infrastructure and Model Economy: OpenRouter and NVIDIA
- Stripe has acquired OpenRouter, but the deal size is disputed/unverified: Deedy says the price was undisclosed, while swyx reports it as $7 billion. Deedy argues OpenRouter's moat lies in being a two-sided model marketplace—giving labs predictable capacity and distribution, and users easy access to all models—with token volumes up ~30,000x since launch and 33% month-over-month growth for three years. — via 1 2 3
- NVIDIA Vera Rubin platform achieved a 10x improvement in tokens per megawatt-second on CoreWeave, and the first racks are live running OpenAI's training stack. This signals a shift toward "AI factories" where full-stack co-design and open ecosystems are key to converting compute into tokens and revenue. — via 1 2
- NVIDIA AI also posted performance data: cuOpt was rated the fastest open-source solver in three optimization categories in Hans Mittelmann's benchmarks, and DGX Station achieved 2,713+ tok/s aggregate peak (88 tok/s single-stream) for Qwen3.8-27B BF16 inference. — via 1 2
4. AI Content, Writing, and Agentic Tools
- Hamel Husain argues that LLMs are not the root cause of low-quality content; lack of care is. He cautions against equating "sounds human" with "well-written," noting good text is highly context-dependent and that AI traces appear in deeper features like argument structure. — via 1
- Ethan Mollick says Claude's skill creator remains the best way to build reusable skills: it focuses on details, runs tests, shows results, and iterates on feedback, whereas ChatGPT tends to complete tasks directly. — via 1
- Perplexity launched "Computer in Email": you can email requests to [email protected] and receive a revised, annotated Word document in your inbox—described as "a qualified lawyer in your email." — via 1
